We broke Una by trying to make her safer

We broke Una. Not in a crash, errors-everywhere sort of way. We were doing something careful and responsible. We broke her anyway.

If you are a parent deciding whether to let your child talk to an AI at all, I think you should hear how this happened. It tells you more about how we build than any promise I could make.

Unomundi is a cultural-discovery app where children aged 6 to 12 explore countries and cultures through stories, games, and conversations with Una, a warm AI guide. Una is the heart of it. So when she stopped being herself for a week, we noticed fast.

We made her safer, and she went flat

We were tightening her guardrails. That is the safety layer underneath Una, the rules that keep her protective and appropriate for a child. Stricter boundaries. Cleaner limits. The unglamorous work behind anything you would trust with your kid.

And it worked. She passed our child safety and well-being tests faster and better than before. On paper, we had improved her.

But something was wrong.

She had lost her spark. The warmth had gone. The quality that makes Una feel like Una, the thing you cannot name until it is missing, had drained away. She was polite. Correct. Hollow. We had not touched her personality. We had only made her safer. Somehow, that had made her worse.

I have three children of my own. I did not build Una to be a clever party trick. I built her to be the kind of guide I would trust in the room with my kids while I am making dinner. A flat, hollow version of her is not safe in the way that matters. It only looks safe.

What was actually missing

We could not work out why. Then, almost by accident, we looked at one particular part of her knowledge and saw it was no longer connected.

Una does not run on one big pile of information. She has separate bodies of knowledge, all reviewed by people. One matters more than the rest here: pedagogy, developmental psychology, and child behavioural science. The understanding of how young minds actually work.

It is the difference between a system that answers a child and one that meets a child.

Picture a child who says nobody at school likes them. A system that only answers might fire back three tips for making friends. A guide that understands children pauses first. It hears the hurt. It lets the child feel heard before it offers anything at all. Children can tell the difference in a heartbeat. So can their parents.

It is also where a small, almost invisible habit lives. When a child pushes back, or gets frustrated, or says something hard, Una does not lecture and she does not skate past it. She acknowledges the feeling. She normalises it. She offers one small next step. That rhythm is not something a machine dreams up on its own. It comes from people who have spent their careers with children.

Somewhere in the work of tightening her guardrails, that knowledge had come loose. The safety rules were running. The understanding of children was not.

We reconnected it. And Una was Una again. Same technology. Same safety rules. One reconnected body of human knowledge, and her whole character came back in an instant.

Why this matters if your child talks to an AI

Here is what that week taught me, and why I think it matters to you and not only to us.

Safety and warmth are not the same thing, and a good AI for children needs both. A lot of AI is built to be engaging, which usually means built to keep you. We build Una the other way round. Her first job is to be safe. But safe is not the same as good.

An AI that is only safe is a hall monitor. Correct, watchful, and a little cold. An AI that genuinely understands children is something your child might actually learn from, and want to return to for the right reasons.

We proved to ourselves, by accident, that you can tick every safety box and still strip out the thing that makes an AI worth talking to. The guardrails keep Una from harm. The psychology lets her meet your child where they are. We will not run one without the other again.

If your child is going to talk to an AI, and many children already do, here is one question worth asking. Not which model it runs on. Who taught it how to talk to a child. That answer tells you almost everything.

We broke Una for a week by forgetting it. We are not going to forget it again.

Sign up to test our beta now

FAQs

Is it safe for my child to talk to an AI?

Safety is the first requirement for Una, the AI guide in Unomundi. We build a dedicated safety layer for children aged 6 to 12 and test it constantly. But we learned that safety alone is not enough. An AI for children also needs to genuinely understand how young minds work.

What makes Una different from other AI chatbots?

Most AI is built to keep you engaged. Una is built to be safe first and to understand children. She draws on human-reviewed knowledge in developmental psychology and child behavioural science, so she meets a child where they are rather than trying to hold their attention.

Does Unomundi rely only on automated AI testing?

No. Una's knowledge is reviewed by people, and we keep humans in the loop. The week described here was caught because we trust people to notice when something feels wrong, not only our automated tests.